Crafting a unique and promising research hypothesis is a critical skill for any scientist. It can also be time-consuming: new PhD candidates may spend the first year of their program trying to decide exactly what to explore in their experiments. What if artificial intelligence could help?
MIT researchers have created a way to autonomously generate and evaluate promising research hypotheses across fields, through human-ai collaboration. In a new paper, they describe how they used this framework to create evidence-based hypotheses that align with unmet research needs in the field of biologically inspired materials.
Published on Wednesday in Advanced materialsThe study was co-authored by Alireza Ghafarollahi, a postdoc in the Laboratory of Atomistic and Molecular Mechanics (LAMM), and Markus Buehler, Jerry McAfee Professor of Engineering in the departments of Civil and Environmental Engineering and of Mechanical Engineering at MIT and director of LAMM.
The framework, which the researchers call SciAgents, consists of multiple ai agents, each with specific capabilities and access to data, that leverage “graphical reasoning” methods, where ai models use a knowledge graph that organizes and defines the relationships between various scientific concepts. The multi-agent approach mimics the way biological systems are organized as groups of elementary building blocks. Buehler points out that this principle of “divide and conquer” is a prominent paradigm in biology at many levels, from materials to insect swarms to civilizations, all examples in which the total intelligence is much greater than the sum of the capabilities of individuals.
“By using multiple ai agents, we try to simulate the process by which communities of scientists make discoveries,” says Buehler. “At MIT, we do this by having a group of people from different backgrounds working together and meeting in coffee shops or in the MIT Infinite Corridor. But that is very coincidental and slow. Our mission is to simulate the discovery process by exploring whether ai systems can be creative and make discoveries.”
Automating good ideas
As recent developments have shown, large language models (LLMs) have demonstrated an impressive ability to answer questions, summarize information, and execute simple tasks. But they are quite limited when it comes to generating new ideas from scratch. The MIT researchers wanted to design a system that would allow ai models to perform a more sophisticated multi-step process that goes beyond remembering information learned during training, to extrapolate and create new knowledge.
The basis of their approach is an ontological knowledge graph, which organizes and establishes connections between various scientific concepts. To make the graphs, researchers feed a set of scientific articles into a generative ai model. In previous work, Buehler used a field of mathematics known as category theory to help the ai model develop abstractions of scientific concepts as graphs, based on defining relationships between components, in a way that could be analyzed by other models. through a process called graph reasoning. . This focuses ai models on developing a more principled way of understanding concepts; it also allows them to generalize better across domains.
“This is really important for us when creating science-focused ai models, since scientific theories are often rooted in generalizable principles and not simply knowledge retrieval,” says Buehler. “By focusing ai models on 'thinking' in such a way, we can go beyond conventional methods and explore more creative uses of ai.”
For the most recent paper, the researchers used about 1,000 scientific studies on biological materials, but Buehler says knowledge graphs could be generated using many more or fewer research articles from any field.
Once the graph was established, the researchers developed an artificial intelligence system for scientific discoveries, with multiple models specialized to perform specific functions in the system. Most components were built from OpenAI's ChatGPT-4 series models and used a technique known as in-context learning, in which prompts provide contextual information about the model's role in the system while also They allow you to learn from the data provided.
The individual agents in the framework interact with each other to collectively solve a complex problem that none of them could solve alone. The first task assigned to them is to generate the research hypothesis. LLM interactions begin after a subgraph has been defined from the knowledge graph, which can occur randomly or by manually entering a couple of keywords analyzed in the articles.
In the framework, a language model the researchers called an “ontologist” is tasked with defining scientific terms in articles and examining the connections between them, fleshing out the knowledge graph. Then, a model called “Scientist 1” crafts a research proposal based on factors such as its ability to discover unexpected and novel properties. The proposal includes a discussion of possible findings, the impact of the research, and a conjecture about the underlying mechanisms of action. A “Scientist 2” model expands on the idea, suggesting specific experimental and simulation approaches and making other improvements. Finally, a “critical” model highlights its strengths and weaknesses and suggests further improvements.
“It's about building a team of experts who don't all think the same way,” Buehler says. “They have to think differently and have different capabilities. The critical agent is deliberately programmed to criticize others, so not everyone agrees and says it's a great idea. You have an agent who says, 'Here's a weakness, can you explain it better?' That makes the result very different from that of the individual models.”
Other agents in the system can search existing literature, which provides the system with a way to not only evaluate the feasibility but also to create and evaluate the novelty of each idea.
Strengthen the system
To validate their approach, Buehler and Ghafarollahi built a knowledge graph based on the words “silk” and “energy-intensive.” Using this framework, the “Scientist 1” model proposed integrating silk with dandelion-based pigments to create biomaterials with improved optical and mechanical properties. The model predicted that the material would be significantly stronger than traditional silk materials and would require less energy to process.
Scientist 2 then made suggestions, such as using specific molecular dynamics simulation tools to explore how the proposed materials would interact, adding that a good application for the material would be a bioinspired adhesive. The Critic model then highlighted several strengths of the proposed material and areas for improvement, such as its scalability, long-term stability, and the environmental impacts of solvent use. To address those concerns, the reviewer suggested conducting pilot studies for process validation and conducting rigorous analyzes of the material's durability.
The researchers also carried out other experiments with randomly chosen keywords, which produced several original hypotheses about more efficient biomimetic microfluidic chips, improving the mechanical properties of collagen-based scaffolds and the interaction between graphene and amyloid fibrils. to create bioelectronic devices.
“The system was able to generate these new, rigorous ideas based on the path from the knowledge graph,” Ghafarollahi says. “In terms of novelty and applicability, the materials seemed robust and novel. “In future work, we will generate thousands, or tens of thousands, of new research ideas, and then we can categorize them, try to better understand how these materials are generated and how they could be improved further.”
In the future, researchers hope to incorporate new tools for retrieving information and running simulations into their frameworks. They can also easily swap out the basic models in their frameworks for more advanced models, allowing the system to adapt to the latest innovations in ai.
“Because of the way these agents interact, an improvement in a model, even if slight, has a huge impact on the overall behaviors and output of the system,” Buehler says.
Since publishing a preprint with open source details of their approach, the researchers have been contacted by hundreds of people interested in using the frameworks in various scientific fields and even areas such as finance and cybersecurity.
“There are a lot of things you can do without having to go to the lab,” Buehler says. “Basically, what you want is to go to the laboratory at the end of the process. The lab is expensive and time-consuming, so you need a system that can drill down to the best ideas, formulate the best hypotheses, and accurately predict emerging behaviors. “Our vision is to make this easy to use, so you can use an app to contribute other ideas or drag in data sets to really challenge the model and make new discoveries.”
